‘AI projects that Google’s India team is working on are super exciting’

From the initial development of Google’s AdSense for Content product to co-founding the Google Brain project, Jeffrey Dean is the driver behind Google’s efforts to democratise Artificial Intelligence. His teams are working on systems for speech recognition, computer vision, language understanding, and various other machine learning tasks. BusinessLine spoke to Dean, Google Senior Fellow and Senior Vice President for Google Research, to get insights into AI projects being done in India,, concerns around data collection and possible misuse of AI. Excerpts.

The vast interpretive computing processing that has happened throughout the history of the computing field has finally got us to the point where machine learning methods can actually teach computers to see, or to hear, or to do speech recognition, more advanced language understanding and translation than ever before. We started to think about how can we use machine learning to take advantage of the fact that computers can now see. Think about the evolutionary times when animals evolved eyes...that was hugely influential I am sure. So now we are at the same point in computing. All of a sudden, computers can open their eyes and see the world around us and that has all kinds of implications for things like healthcare and education.

I have looked at different sets of projects that the research team at the Bengaluru office has been working on. They are all super exciting and very different. For example, there is an organisation called Armaan which offers free mobile voice call service to expecting mothers. However, over 50 per cent of women drop off from the programme. We partnered with IIT-Madras to help them use AI to predict the risk of expectant mothers dropping out of healthcare programmes. They successfully tested the results of our algorithm with a randomised control trial with 6,000+ beneficiaries and noticed a 13% improvement in retention. They are now expanding this AI approach to benefit 2 million plus women, and have further plans to scale this AI technology to 18 million plus women. Another one that I am pretty excited about is “bolo” (now called Read Along). We first tested this in India and are now rolling out all around the world. It reads a story and listens to the child narrate the story back and uses speech recognition and machine learning to coach the student on how they can pronounce words better.

These are the kinds of things that are outlined in our company-wide AI principles. In healthcare, it is especially important to be cognizant of these kinds of issues because .you are making very consequential decisions about people’s health. If you take the case of diabetic retinopathy in India, first of all there is a pretty large number of people who should be screened every year for diabetic retinopathy. Since it is a degenerative eye disease, if you catch it in time, it is very treatable. And in India, there is a shortage of more than 100,000 eye doctors to do even routine screening. But a machine learning model can help do the diagnostic screening that frees up ophthalmologists to spend more of their time treating the patients that the system has identified as those that need treatment.

You want these systems to be able to operate even in a situation where they do not have good connectivity. For example,, the speech recognition system on my phone works even in airplane mode. It can do better when there is internet connectivity available and it can communicate with the dataset to have a more powerful speech recognition model, but it works fine without that. The Read Along application actually runs on very low-end smartphones that is accessible to many income levels in rural India. It does all of its processing on the device itself. Through the advancements in the semiconductor industry, mobile phones are becoming more affordable and more capable. One of the things that can help run more machine learning models on-device is having customised processing elements in the phone itself that can speed up AI-related application and run your computer vision algorithm and the speech recognition algorithms at very high level of accuracy but with a very low power consumption.

We are doing a lot of research to enable new kinds of machine learning algorithms to be able to learn from less data than other algorithms. The second thing I would say is that we are guided by our AI principles in how we think about using AI and machine learning. And so like any technology, it can be put to really amazing uses, improving health care and education, but can also be put to very negative uses. And, I think it is up to us and the society to really help shape how some of these technologies are used. The third thing I would say is, we have quite a lot of focus on how we can learn from data in a privacy-preserving way.

Technologies change quickly, and policies and laws appropriately change more slowly. What you want as a citizen is for these policies by governments to be made in an informed way. I have no problem with policy makers looking at a situation and understanding the nuances of it and making good decisions that are good for the citizens of the country. I just want those decisions to be made in an informed and helpful way.

Also read:

Also read:

COMMENT NOW